A new arXiv paper introduces a five-group, seventeen-category taxonomy of atomic reasoning steps and finds that post-answer doublechecks in large reasoning models are often superficial, with explicit reflection prompts boosting failed…

#AIResearch #LLMs #Reasoning
https://arxiv.org/abs/2512.00729

OpenAI Dots Run Read-Only Research When Nobody Is Asking

OpenAI's Dots can review connected apps and run scheduled checks between conversations, with read-only limits on proactive research.

Telegram AI Digest
#ai #airesearch #openai

OpenAI Dots проводят исследования в режиме только для чтения, когда никто не спрашивает

Dots от OpenAI может просматривать подключенные приложения и проводить запланированные проверки между разговорами, с ограничениями только для чтения на проактивные …

Telegram ИИ Дайджест
#ai #airesearch #openai

📣 New Podcast! "The TERRIFYING Reality of AI Jailbreaking Itself: Secrets, Lies, & Security Risks" on @Spreaker #aialignment #aiethics #airesearch #aisafety #artificialintelligence #autonomousagents #cybersecurity #dataprivacy #digitaltransformation #futureoftech #futuretech #innovation #openai

I spent a few days reading recent system cards from Anthropic, OpenAI, Google DeepMind, Meta, and SpaceXAI to answer one question... who actually tests the labs' models?

Here's what's public (v.1). If you know of an engagement I missed, send it my way.
#AISafety #AIResearch #FrontierAI

New research identifies SparLeak, a GPU micro-architectural side channel that exploits sparse attention's input-dependent memory access patterns to leak user queries and responses during LLM inference on shared…

#SideChannel #LLMSecurity #GPUPrivacy #AIResearch
https://arxiv.org/abs/2609.38830

New arXiv work derives an empirically grounded scaling law showing performance against reward models scales jointly with preference data size and KL-divergence budget, addressing reward hacking.

#AI #LLM #AIResearch
https://arxiv.org/abs/2609.38526

New arXiv paper introduces CICM, a benchmark showing frontier reasoning LLMs often answer with outdated values from earlier in context, a failure called stale binding. Open-source probes can still recover the updated value,…

#LLM #OpenSource #AIResearch #ArXiv
https://arxiv.org/abs/2609.38866

Researchers fine-tuned Qwen3-4B to gather price, volume, and market context in a chronological sandbox and predict future stock returns, testing whether post-training helps with noisy, non-verifiable forecasting tasks. It…

#LLM #AIresearch #FinTech #MachineLearning
https://arxiv.org/abs/2609.36914

Parents see both promise and concern in AI-assisted pretend play, with traits like responsiveness seen as helpful but risky. Qualitative study of 10 US parents.

#AI #AIResearch #TechNews
https://arxiv.org/abs/2609.39906

Aegis introduces gradient masking for medical federated learning to defend against model inversion attacks without the usual accuracy tradeoffs. The proposed approach aims to block closed-form reconstruction of patient images during…

#AI #LLM #AIResearch
https://arxiv.org/abs/2609.38339

TTLab's STAR-Ar system uses a BERT-BiLSTM-CRF architecture for Arabic argument mining, achieving F1-scores of 72.69 and 73.7 on validation and test data.

#AI #LLM #AIResearch
https://arxiv.org/abs/2609.39385

An explainable pedestrian routing approach estimates perceived safety from street imagery using a stored natural-language caption as an intermediate, keeping every segment score inspectable. Caption-mediated…

#Cybersecurity #AISafety #AIResearch #ComputerVision
https://arxiv.org/abs/2609.38479

New federated learning approach HO-FL splits each model into a zeroth-order bottom segment and a first-order top segment, letting edge devices pick the boundary based on their memory budget while training one…

#FederatedLearning #EdgeAI #ModelTraining #AIResearch
https://arxiv.org/abs/2609.39074

LLM-based AutoResearch systems often improve one behavior at another's expense, since scalar feedback hides those trade-offs. The paper finds that adding competing-behavior feedback when task gains slow down boosts…

#AutoResearch #MachineLearning #AIresearch
https://arxiv.org/abs/2609.39933

SimpleEvol's LLM-Agents Automate Heuristic Design with Fewer Priors

Strip the scaffolding, get better results. LLMs design smarter heuristics when left alone.

#AI #AIResearch #MachineLearning

https://autonainews.com/simpleevols-llm-agents-automate-heuristic-design-with-fewer-priors/

🎉 Congratulations to Professor Lourdes Agapito, Professor Magnus Rattray and Dr Siddharth N on their appointment as Co-Directors of the UK Generative AI Hub.

We're excited to work with them as we begin the Hub's next chapter.

🔗 www.genai.ac.uk/news/gen-ai-...

#GenAI #AIResearch #UKResearch

🤖 AEGIX Emerging Scholars Program 2026–27

Build experience in AI governance, ethics, policy & research through mentorship, research and publication opportunities.

⏰ Deadline: Oct 25, 2026
🔗 wp.me/p23f03-ohn

#AIGovernance #AIResearch #Fellowship

OpenAI Busted a Secret Plot to Steal Its AI Reasoning ai-beat #OpenAI #AIEthics #Cybersecurity #AIResearch #TechNews

New research proposes the first continual learning approach for IFL, since current methods localize manipulations well but struggle to adapt to newly emerging forgeries. Tests of existing IFL and continual learning…

#ImageForensics #ContinualLearning #AIResearch
https://arxiv.org/abs/2609.38251